RF-DETR Sets COCO Real-Time Detection SOTA at ICLR 2026, Apache 2.0
Roboflow published RF-DETR, a DINOv2-backbone transformer detector accepted at ICLR 2026 that sets a new state of the art on COCO real-time object detection. The Apache 2.0-licensed RF-DETR-L achieves 56.5 AP-50:95 at 6.8ms on a T4 GPU—outperforming YOLO26-L (54.1 AP, 5.7ms) on accuracy. The largest variant, RF-DETR-2XL, hits 60.1 AP at 17.2ms, beating every YOLO11 and YOLO26 size class on accuracy. Crucially, RF-DETR is Apache 2.0 up to the L size, removing the AGPL-3.0 viral license constraint that has blocked production deployment of YOLO models.
Why It Matters
Apache 2.0 licensing on a COCO-SOTA real-time detector is a category unlock: production computer vision applications that previously used YOLO under commercial exception now have a superior, commercially clean alternative with no copyleft risk.